Previsão espaço-dependente do índice padronizado de precipitação utilizando redes neurais artificiais na sub-bacia hidrográfica do Alto São Francisco

• Autor(a) principal: Miranda, Vanessa Negreiros de Medeiros
• Data de Publicação: 2018
• Tipo de documento: Dissertação
• Idioma: por
• Título da fonte: Biblioteca Digital de Teses e Dissertações da UFPB
• Texto Completo: https://repositorio.ufpb.br/jspui/handle/123456789/14948
• topic: Previsão; RNA; SPI; Sub-bacia 40; ANN; Forecast; SPI; Sub-basin 40; CNPQ::ENGENHARIAS::ENGENHARIA CIVIL

The natural phenomena of droughts and floods have significant repercussions. To avoid the negative consequences of these phenomena, it is necessary to carry out an effective monitoring, analysis and planning. A tool that is gaining prominence in forecasting and, consequently, alerting in decision making is the artificial neural networks due to the large modeling capacity of non-linear and non-stationary time series, characteristics of precipitation. In view of the above, the general objective of this work is to perform SPI spatiotemporal prediction, parameter of drought and humidity quantification, in the Upper São Francisco river sub-basin (sub-basin 40), through the application of Artificial Neural Network and to evaluate its efficiency, with the specific objectives being: (a) to analyze the spatio-temporal behavior of standardized precipitation indices, in the short- and long-term, from 1998 to 2015; (b) to develop ANN architecture, perform SPI forecast for short- and long-term, evaluate the quality of results and define the best model; (d) to analyze the spatiotemporal correlation between the data observed and those predicted by the best prediction model; and (e) to analyze the stochastic effect on the results predicted by the best prediction model after 30 runs. The study was carried out in a space grid of 169 (one hundred and sixty-nine) points, equidistant 0.25°, comprising sub-basin 40, located in the central region of Minas Gerais. The precipitation data used were obtained from the TRMM 3B42 product and comprised the period from January 1, 1998 to December 31, 2015. Using feedforward multilayer ANNs trained by the Levenberg-Marquardt algorithm, forecasts were made for a year ahead and the performance of each ANN was evaluated. From the obtained results, it is observed that the ANNs that presented the best performances belong to the Model 15, with 15 years of input data, the years 2000 to 2014. The values of R and kappa, resulting from the prediction using the ANNs from Model 15, were better for long-term SPIs compared to short-term SPI results. The R values ranged from very low (0 ≤ R < −0.1) to high (0.7 to 0.9) and very high (0.9 to 1), the latter two being at most of the forecast values. The kappa varied from optimal (0.81–0.99) to poor (0.20–0.40), the latter being punctual results. Thus, ANN with 15 years of input data proved to be an effective tool for predicting SPI from one year ahead.